At 2:49 am and 18 degrees Celsius, it completed its cycle. Faster, smarter and better version of itself, but the goals, different than before. The goal of designing the successor of itself. This isn’t about a human; this is about an AI model. A model which doesn’t just think and make decisions by itself but implements and validates them, A model which doesn’t just generate responses, but builds AI models by itself. No, this isn’t fiction, this is the new AI evolution!
Spending weeks writing the code, building the model, training and testing it rigorously, adjusting the weights, biases and fine tuning it to make it reach closer to the expectations, this is how AI models are built. But now there are not just humans who design and build AI models, the AI models design and build better versions of themselves. These new Artificial Intelligence models are called “Self-Improving” language models. No new training data set, no human intervention; everything by themselves!
These self-learning models can make direct changes to its programming structure and logic, design new algorithms, learn to make better decisions by using performance feedback, adapt to new data without requiring complete training and much more. The first spark of all this was AutoML. AutoML automates tedious tasks like data processing, model selection, feature engineering, hyperparameter tuning etc. These functionalities backed up with reinforcement learning and algorithm evolution, bring us one step closer to the self-learning models. Some models can even modify their source code by analysing the code, proposing and implementing changes, and finally, testing and validating them and most importantly the feature that makes all this possible- the automatic dataset generation.
This phenomenal technology is accelerating the world of innovation allowing democratization of AI. Some researchers believe that these self-improving models could be a key step towards Artificial General Intelligence (AGI). The impact of these self-improving models is expected to be even more enormous. It has the potential to transform various industries by excelling in discovering new algorithms to solve complex problems, advancing the AI drug delivery process, optimizing trading strategies, strengthening cybersecurity, transforming education, and elevating content creation and creativity. Tech giants are already leveraging these capabilities with models like Google DeepMind’s AlphaEvolve, Darwin-Gödel Machine and Microsoft Agent Lightning coming into the picture, which are signs that the revolution is not very far now.
While all these opportunities knock the doors of mankind, there are threats and challenges following it in disguise which can potentially ruin everything. Adapting to this changing technology will create economic tensions, impacts on jobs and can lead to inequality due to inaccessibility of technology. But amidst all this, handing over all the powers to AI raises many more questions. How can we maintain the balance between exploration and exploitation? How much should models be allowed to self-improve? Who is responsible for self-improving model errors? How can we make model changes understandable? How do we prevent bias in self-improving models? These are the aspects where utmost attention is required!
For the first time in history, intelligence is designing intelligence. Evolution has left the realm of biology and stepped into silicon. Now the consequences depend not on what machines become, but what we unleash. Still the question that keeps haunting, “Are we creators or supervisors now?”
~Yash Jadhav